Title Fixed Point Theory in Weak Second-Order Arithmetic
Author(s) SHIOJI, NAOKI; TANAKA, KAZUYUKI
Citation 数理解析研究所講究録 (1988), 669: 44-77
Issue Date 1988-08
URL http://hdl.handle.net/2433/100734
Right
Type Departmental Bulletin Paper
Textversion publisher
Kyoto University
44
Fixed Point Theory
in
Weak Second-Order Arithmetic
NAOKI SHIOJI AND KAZUYUKI TANAKA(塩路 直樹) (田中 一之)
Department of Information Science
Tokyo Institute of Technology
Ookayama, Meguroku, Tokyo 152, Japan
1. Introduction
The purpose of this paper is to develop part of functional analysis concerned with
fixed point theorems within a relatively weak subsystem of second-order arithmetic,
known as $WKL_{0}$ . The interest of $WKL_{0}$ has been well established through ongoing
program, called Reverse Mathematics, whose ultimate goal is to answer the following
question: What set existence axioms are needed to prove the theorems of ordinary
mathematics? For information on the program, see [6], [7], [15], [16].
We here briefly describe the system $WKL_{0}$ and two other related systems $RCA_{0}$ ,
$ACA_{0}$ . $RCA_{0}$ is the system of recursive comprehension and $\Sigma_{1}^{0}$ induction. This is the
weakest system we shall consider, but is still strong enough to develop some basic
theory of continuous functions and countable algebras. The system $WKL_{0}$ consists
/
数理解析研究所講究録第 669巻 1988年 44-77
45
of $RCA_{0}$ plus the weak K\"onig’s lemma: every infinite tree of sequences of $0s$ and
l’s has an infinite path. The first-order part of $WKL_{0}$ is the same as that of $RCA_{0}$ ,
but $WKL_{0}$ proves many important theorems which are not provable in $RCA_{0}$ , e.g.,
the Heine-Borel theorem. From the viewpoint of the traditional proof theory, both
$R_{I}CA_{0}$ and $WKL_{0}$ are as weak as Primitive Recursive Arithmetic, which is regarded
as a formal system suited for Hilbert’s finitism to a great extent. The third system
$ACA_{0}$ consists of $RCA_{0}$ plus the arithmetical comprehension axiom. This system is
strictly stronger than $WKL_{0}$ , and its first-order part is just Peano arithmetic.
In this paper, we discuss several forms of fixed point theorems and their applica-
tions within $WKL_{0}$ . A topological space $X$ is said to possess the fixed point property
if every continuous function $f$ : $Xarrow X$ has a point $x\in X$ such that $f(x)=x$ . An
elementary argument in $RCA_{0}$ shows that the unit interval $[0,1]$ has the fixed point
property. However, it is not provable within $RCA_{0}$ that $[0,1]^{2}$ (or the closed unit disc)
has the fixed point property. We indeed show that this statement is equivalent to
$WKL_{0}$ over $RCA_{0}$ . A general assertion of Brouwer’s theorem is that every nonempty
compact convex subset $C$ of $R^{n}$ has the fixed point property. We prove this asser-
tion within $WKL_{0}$ , provided that the set $C$ can be expressed as the completion of a
countable subset of $Q^{n}$ as well as that the complement of $C$ is expressed as the union
of (a sequence of) basic open sets. In fact, the assertion with the same proviso can
be proved in $RCA_{0}$ , since if the compact set $C$ contains two or more points, the line
segment connecting two distinct points in $C$ must be compact, which implies $WKL_{0}$ ,
and otherwise the assertion is trivial. We then extend Brouwer’s theorem to its infi-
nite dimensional analogue (the Tychonoff-Scahuder theorem for $R^{N}$) by adapting Ky
Fan’s technique based on the Knaster-Kuratowski-Mazurkiewicz theorem (see [5]).
2
46
As an application of the fixed point theorem for $R^{N}$ , we prove the Cauchy-Peano
theorem for ordinary differential equations within $WKL_{0}$ , which was first shown by
Simpson [14] without reference to the fixed point theorem.
We next discuss the Markov-Kakutani fixed point theorem which asserts the ex-
istence of a common fixed point for certain families of affine mappings. While the
original proof due to Markov depended on TychonofF’s theorem, Kakutani [11] gave
a direct proof. We adapt Kakutani’s proof for $RCA_{0}$ . Kakutani [11] also proved
that the Markov-Kakutani theorem implies the Hahn-Banach theorem. We use his
technique to reprove the Hahn-Banach theorem for separable Banach spaces within
$WKL_{0}$ , which was first shown in a direct but somewhat unnatural way by Brown and
Simpson [3].
In section 2, we define the formal systems $RCA_{0},$ $WKL_{0}$ and $ACA_{0}$ . Section 3 is
devoted to the development of basic concepts of real analysis. In section 4, we discuss
uniform continuity and integration. In section 5, we investigate what set existence
axioms are needed to prove some variants of Brouwer’s fixed point theorems. In
section 6, we prove, within $WKL_{0}$ , the Tychnoff-Schauder theorem for $R^{N}$ , and apply
it to the Cauchy-Peano theorem. In section 7, we prove, within $WKL_{0}$ , the Markov-
Kakutani fixed point theorem for $R^{N}$ , and then use it to prove the Hahn-Banach
theorem for separable Banach spaces within $WKL_{0}$ .
2. The systems $RCA_{0},$ $WKL_{0}$ and $ACA_{0}$
In this section, we describe the formal systems $RCA_{0},$ $WKL_{0}$ , and $ACA_{0}$ , following
[3], [16]. The reader who is familiar with these systems may skip to the next section.
The language of second-order arithmetic is a two-sorted language with number
3
47
variables $i,j,$ $k,$ $l,$ $m,$ $\ldots$ and set variables $X,$ $Y,$ $Z,$ $\ldots$ . Numerical terms are built up
from number variables and constant symbols $0$ and 1 by means of binary operations
$+and\cdot$ . Atomic formulas are $t_{1}=t_{2},$ $t_{1}<t_{2}$ , and $t_{1}\in X$ , where $t_{1}$ and $t_{2}$ are numer-
ical terms. Formulas are built up from atomic formulas by means of propositional
connectives, number quantifiers $\forall n$ and $\exists n$ , and set quantifiers $\forall X$ and $\exists X$ .
A formula is said to be arithmetical if it contains no set quantifiers. An arithmeti-
cal formula is said to be $\Sigma_{0}^{0}$ if all of its number quantifiers are bounded, i.e., of the
form $\forall n(n<tarrow\ldots)$ or $\exists n(n<t\ . . .)$ . An arithmetical formula is said to be $\Sigma_{1}^{0}$
(resp. $\Pi_{1}^{0}$ ) if it is of the form $\exists m\phi(m)$ (resp. $\forall m\phi(m)$ ) where $\phi(m)$ is a $\Sigma_{0}^{0}$ formula.
The system $RCA_{0}$ consists of the ordered semiring axioms for $(N, +, \cdot, 0,1, <)$
together with the scheme of $\triangle_{1}^{0}$ comprehension and $\Sigma_{1}^{0}$ induction. The $\triangle_{1}^{0}$ (recursive)
comprehension scheme consists of all formulas of the form
$\forall n(\phi(n)rightarrow\psi(n))arrow\exists X\forall n(n\in Xrightarrow\phi(n))$ ,
where $\phi(n)$ is a $\Sigma_{1}^{0}$ formula, $\psi(n)$ is a $\Pi_{1}^{0}$ formula, and $X$ does not occur freely in
$\phi(n)$ . The $\Sigma_{1}^{0}$ induction scheme consists of all formulas of the form
$\phi(0)\wedge\forall(\phi(n)arrow\phi(n+1))arrow\forall n\phi(n)$ ,
where $\phi(n)$ is a $\Sigma_{1}^{0}$ formula. At all times, we assume the law of the excluded middle.
The system $ACA_{0}$ consists of $RCA_{0}$ plus the arithmetical comprehension scheme
$\exists X\forall n(n\in Xrightarrow\phi(n))$ ,
where $\phi(n)$ is arithmetical and $X^{-}$ does not occur freely in $\phi(n)$ . We can easily see
that any arithmetical instance of induction scheme is provable in $ACA_{0}$ , and that
$ACA_{0}$ is a conservative extension of first-order Peano arithmetic.
$\not\simeq$
48
The system $WKL_{0}$ is intermediate between $RCA_{0}$ and $ACA_{0}$ . Within $RCA_{0}$ , we
define $Seq_{2}$ to be the set of (codes for) finite sequences of $0’ s$ and l’s. A set $T\subseteq Seq_{2}$
is said to be a tree if any initial segment of a sequence in $T$ is also in $T$ . A path
through $T$ is a tree $P\subseteq T$ such that for any two sequences in $P$ , one of them is
an initial segment of the other. The axioms of $WKL_{0}$ are those of $RCA_{0}$ plus weak
K\"onig’s lemma: every infinite $tree\subseteq Seq_{2}$ has an infinite path. $WKL_{0}$ is known to
be conservative over Primitive Recursive Arithmetic with respect to $\Pi_{2}^{0}$ sentences.
3. The basic concepts of real analysis
This section is devoted to the development of basic concepts of real analysis,
including continuous functions, compactness and convexity in the spaces $R^{n}$ and $R^{N}$ ,
all within the system $RCA_{0}$ .
First of all, we use the symbol $N$ informally to denote the set of natural numbers.
We introduce total functions from $N$ into $N$ by encoding them as sets of ordered
pairs. Within $RCA_{0}$ , we can define most of elementary numerical functions (e.g.,
the exponential function $m^{n}$ ) in the usual way, and can prove their basic properties.
We then define (codes for) rational numbers to be certain ordered pairs of natural
numbers. The arithmetical operations on the rational numbers are defined in the
standard way. We write $Q$ for the set (or the field) of rational numbers thus defined.
We define an (infinite) sequence of rational numbers to be a function $f$ : $Narrow Q$ ,
and denote such a sequence by $\{a_{n} : n\in N\}$ or simply by $\{a_{n}\}$ , where $a$ . $=f(n)$ . $A$
real number is defined to be a sequence $\{a_{n}\}$ of rational numbers such that $VnVi(|a.-$
$a_{n+i}|\leq 2^{-n})$ . We use $R$ informally to denote the set of all real numbers. Note that
$R$ does not formally exist as a set within $RCA_{0}$ . Two real numbers $\{a_{n}\}$ and $\{b_{n}\}$ are
$6^{-}$
49
defined to be equal, $\{a_{n}\}=\langle b_{n}$ }, iff $\forall n(|a_{n}-b_{n}|\leq 2^{-n+1})$ . The relation $<$ is defined
by { $a_{n}\rangle$ $<\{b_{r;}\rangle$ iff $\exists n(b_{n}-a_{n}>2^{-n+1})$ . Then it is easy to see that for any two real
numbers { $a_{n}$ ) and $\{b_{n}\}$ , exactly one of the following holds (in $RCA_{0}$ ) $:\{a_{n}\}<.$ $b_{n}$ ),
$\{a_{n}\}=\langle b_{n}$ }, \langle $a_{n}$ } $>\langle b_{n}$ }. We let $0=\langle 0;n\in N$ } and $1=\langle 1$ : $n\in N$ }. The operations
$+and$ . are defined by
$\{a_{n}\}+(b_{n}\}=\langle a_{n+1}+b_{n+1}\}$ ,
$\{a_{n})\cdot\{b_{n})=\{a_{n+m}\cdot b_{n+m}\rangle$ ,
where $m$ is the least number such that $\max(|a_{0}|, |b_{0}|)+1\leq 2^{m-1}$ . Within $RCA_{0}$ , one
can prove that $(R, +, \cdot, 0,1, <)$ is an Archimedean ordered field.
An infinite sequence of real numbers is defined to be a doubly indexed sequence
of rational numbers { $a_{mn}$ : $m,$ $n\in N\rangle$ such that for each $m,$ { $a_{mn}$ : $n\in N\rangle$ is a
real number. Such a sequence of real numbers is also denoted { $x_{m}$ : $m\in N\rangle$ , where
$x_{m}=\langle a_{mn}$ : $n\in N$ }. We write $R^{N}$ for the set of infinite sequences of real numbers.
Similarly, an n-tuple (or finite sequence with length n) of real numbers, for $n\geq 1$ , is a
doubly indexed sequence of rational numbers \langle $a_{ij}$ : $i<n,j\in N$} such that for each
$i<n,$ { $a_{ij}$ : $j\in N\rangle$ is a real number. We write $R^{n}(n\geq 1)$ for the set of n-tuples of
real numbers. In case $n=1,$ $R^{n}$ is identified with $R$ . Two elements in $R^{n}$ (or $R^{N}$ )
are defined to be equal, if their corresponding components are equal with respect to
the equality of $R$ .
We define $\max:R^{n}arrow R$ and $\min:R^{n}arrow R$ as follows:
$\max(\langle\langle a_{ij} : j\in N\} : i<n\})$ $=$ $\{\max\{a_{ij} : i<n\} : j\in N\}$ ,
$\min(\{\{a_{ij} : j\in N)$ : $i<n\rangle$ ) $=$ {$\min\{a_{ij} : i<n\}:j\in N$ ).
It is easy to check that the sequences on the right sides are “real numbers”. The
6
50
finite sum $\sum$ : $R^{n}arrow R$ is defined by
$\sum(\{\{a_{ij} : j\in N\rangle : i<n\rangle)=\langle\sum_{i<n}a_{i(j+n-1)}$ : $j\in N$}.
Its well definedness is also clear. Although $\max,$ $\min$ and $\sum$ could be defined as
continuous functions from $R^{n}$ to $R$ (the notion of continuous functions will be given
later), we treat them like operations such $as+and$ . on $R$ .
The vector addition and scalar multiplication on $R^{n}$ and $R^{N}$ are defined in obvious
way. We define the norm $\Vert\cdot\Vert_{n}$ on $R^{n}$ by I { $x_{i}$ : $i<n\rangle$ $\Vert_{n}=\max|x_{i}|i<n$ Thus $R^{n}$ can be
viewed as a normed linear space. We also use $\Vert\cdot\Vert_{n}$ as a seminorm on $R^{N}$ by letting
$\Vert\{x_{i} : i\in N\rangle\Vert_{n}=\Vert\langle x_{i} : i<n\}\Vert_{n}$ . Then $R^{N}$ is a linear space with countably many
seminorms, and indeed a separable Frechet space.
We next discuss the topology on $R^{n}$ and $R^{N}$ . Let $Q^{n}(n\geq 1)$ be the set of (codes
for) finite sequences of rational numbers with length $n$ . $Q^{n}$ may be regarded as a
subset of $R^{n}$ . We assume that $Q^{n}$ and $Q^{m}$ are disjoint if $n\neq m$ . We then put
$Q^{<N}= \bigcup_{n\geq 1}Q^{n}$ . A code for a basic open set $B_{r}(a)$ in $R^{n}$ (resp. $R^{N}$ ) is an ordered
pair $(a, r)$ with $a\in Q^{n}$ (resp. $Q^{<N}$ ) and $r\in\{0\}\cup Q^{+}$ (the positive rationals). For
$a\in Q^{<N}$ , we define $\dim(a)$ to be the dimension of $a$ , i.e., $a\in Q^{\dim(a)}$ . A point
$x\in R^{n}$ (resp. $R^{N}$ ) is said to belong to a basic open set $B_{r}(a)$ in $R^{n}$ (resp. $R^{N}$),
denoted $x\in B_{r}(a)$ , if $\Vert x-a\Vert_{n}<r$ (resp. $\Vert x-a\Vert_{\dim(a)}<r$ ). By $x\in\overline{B_{r}(a)}$, we
mean $\Vert x-a\Vert_{\dim(a)}\leq r$ . We write $B_{r}(a)\subseteq B_{s}(b)$ in $R^{n}$ (resp. $R^{N}$) to mean that
$\Vert b-a||_{n}+r\leq s$ (resp. dim(b) $\leq\dim(a)$ and I $b-a\Vert_{\dim(b)}+r\leq s$). A basic open set
$B_{r}(a)$ is said to be nonempty if $r>0$ .
A code for an open set $U$ in $R^{n}$ (resp. $R^{N}$ ) is a sequence of (codes for) basic open
sets in $R^{n}$ (resp. $R^{N}$), i.e., a function $\phi$ : $Narrow Q^{n}\cross Q$ (resp. $\phi$ : $Narrow Q^{<N}\cross Q$ )
such that for each $n\in N,$ $\phi(n)$ is a code for a basic open set in $R^{n}$ (resp. $R^{N}$). A
7
51
point $x\in R^{n}$ (resp. $R^{N}$ ) is said to belong to an open set $U$ with code $\phi$ in $R^{n}$ (resp.
$R^{N})$ , denoted $x\in U$ , if it belongs to a basic open set in the sequence, $i.e.$ , there exists
$n\in N$ such that $x\in\phi(n)$ . This definition of open sets is not identical with the
corresponding definition in [3]. Our definition guarantees that for any $x\in R^{n}$ and
any $r\in R$ with $r\geq 0,$ $\{y\in R^{n} : \Vert x-y\Vert_{n}<r\}$ is an open set.
We define closed sets to bejust complements of open sets. Remark that a closed set
has only negative information on its members, and so, in general, it is difficult to deal
with points in a closed set. We thus introduce the notion of countably representable
closed sets, which have both positive and negative information on their members. Let
$f$ be a function from $N$ to $R^{n}$ (or $R^{N}$). We informally identify $f$ with its range, say
$A$ , although $A$ may not exist as a set. A closed set $C\subseteq R^{n}$ (resp. $C\subseteq R^{N}$ ) is said to
be countably represented by $A\subseteq R^{n}$ (resp. $A\subseteq R^{N}$ ) denoted $C=\hat{A}$ , if for all $x\in R^{n}$
(resp. $x\in R^{N}$ ),
$x\in C$ $\Leftrightarrow$ there exists an infinite sequence \langle $a_{i}$ } from $A$ such that
for each $i,$ $\Vert x-a_{i}\Vert_{n}\leq 2^{-i}$ (resp. $\Vert x-a_{i}\Vert_{i+1}\leq 2^{-i}$ ).
For example, $[0,1]^{n}$ is countably represented by $A_{n}=\{\{q_{0},$$\ldots,$
$q_{n-1}\rangle$ $\in Q^{n}$ : $0\leq q_{i}\leq 1$
for each $i<n$}, and $[0,1]^{N}$ by {{ $q_{0},$ $q_{1},$ $\ldots$ , $q_{n},$ $0,0,$ $\ldots\rangle$ $\in Q^{N}$ : $0\leq q_{i}\leq 1$ for each
$i\leq n\}$ .
A code for a continuous partial function $f$ : $R^{n}arrow R^{m}$ (or $R^{N}arrow R^{N}$ ) is a sequence
$\Phi$ of pairs of nonempty basic open sets such that
(i) $(B_{r}(a),\dot{B}_{s_{1}}(b_{1}))\in\Phi\ (B_{r}(a), B_{s_{2}}(b_{2}))\in\Phiarrow\Vert b_{1}-b_{2}\Vert_{k}<s_{1}+s_{2}$ ,
where $k= \min(\dim(b_{1}), \dim(b_{2}))$ ,
(ii) $(B_{r}(a), B_{s_{1}}(b_{1}))\in\Phi\ B_{s_{1}}(b_{1})\subseteq B_{s_{2}}(b_{2})arrow(B_{r}(a), B_{s_{2}}(b_{2}))\in\Phi$ ,
$S$
$(iii)(B_{r_{1}}(a_{1}), B_{s}(b))\in\Phi\ B_{r_{2}}(a_{2})\subseteq B_{r_{1}}(a_{1})arrow(B_{r_{2}}(a_{2}), B_{s}(b))\in\Phi$ .
Although $\Phi$ is formally a function from $N$ to $Q^{n}\cross Q^{+}\cross Q^{m}\cross Q^{+}$ (resp. $Q^{<N}\cross$
$Q^{+}\cross Q^{<N}\cross Q^{+})$ , we write $(B_{r}(a), B_{s}(b))\in\Phi$ if $\Phi(n)=(a, r, b, s)$ for some $n\in N$ .
Intuitively, $(B_{r}(a), B_{s}(b))\in\Phi$ means that $f(B_{r}(a))\subseteq\overline{B_{s}(b)}$ where $f$ is the continuous
partial function encoded by $\Phi$ . A point $x\in R^{n}$ (resp. $R^{N}$ ) is said to belong to the
domain of function $f$ with code $\Phi$ if for all $\epsilon>0$ (resp. for all $\epsilon>0$ and all $i\in N$ ),
there exists $(B_{r}(a), B_{s}(b))\in\Phi$ such that $x\in B_{r}(a)$ and $s<\epsilon$ (resp. $s<\epsilon$ and
$\dim(b)\geq i+1)$ . If a point $x\in R^{n}$ (resp. $R^{N}$ ) is in the domain of $f$ , we define $f(x)$
to be the point $y\in R^{n}$ (resp. $R^{N}$ ) such that if $x\in B_{r}(a)\ (B_{r}(a), B_{s}(b))\in\Phi$ then
$y\in\overline{B_{s}(b)}$. We can prove, within $RCA_{0}$ , that such a $y$ exists uniquely (up to the
equality of points). A continuous partial function $f$ : $R^{n}arrow R^{m}$ or $f$ : $R^{N}arrow R^{N}$ is
said to be a continuous function from a closed set $C$ to a closed set $D$ , if the domain
of $f$ includes $C$ and for each $x\in C,$ $f(x)\in D$ .
In the rest of this section, we discuss the notions of compactness and convexity,
which play the most important roles in fixed point theory. Let $C$ be a closed set in
$1$
$1$
$1$
$1$
$1$
!$i$
$!$
$1$
$1$
;
$!$
$!$
$1$
$1$
$1$
$R^{n}$ or $R^{N}$ . $C$ is said to be compact (in the sense of Heine-Borel) iff for any open set $|$
\langle $B_{i}$ : $i\in N$ } with $B_{i}$ basic open, if $\{B_{i} : i\in N\}$ covers $C$ (i.e., all points in $C$ belong$|$
to some $B_{i}$ ) then there exists $n\in N$ such that $\langle B_{i} : i<n\rangle$ covers $C$ . The notion
of compactness should be distinguished from related concepts such as the Bolzano-
Wierstrass property: every bounded sequence has a limit point. H. Friedman [7] has
shown that $WKL_{0}$ and the compactness (in the sense of Heine-Borel) of $[0,1]$ are
$J$
equivalent over $RCA_{0}$ , and that $ACA_{0}$ and the Bolzano-Wierstrass property of $[0,1]$
are equivalent over $RCA_{0}$ .
We here state the following lemma without proof.
$\eta$
53
3.1 LEMMA$(RCA_{0})$ . The following are pairwise equivalent:
(i) $WKL_{0}$ ,
$(ii)_{n}[0,1]^{n}\subseteq R^{n}$ is compact, $n\geq 1$ ,
(iii) $[0,1]^{N}$ is compact.
For the proof, see Lemma 2.4 and 3.3 of Simpson [14] and the references given there.
We finally discuss the notion of convexity. Let $C$ be a closed set in $R^{n}$ or $R^{N}$ . We
define $C$ to be convex iff for all $x,$ $y$ in $C$ and for all $q\in[0,1]\cap Q,$ $qx+(1-q)y\in C$ .
Then the following holds.
3.2 LEMMA$(RCA_{0})$ . Let $C$ be a convex closed set in $R^{n}$ or $R^{N}$ . Then for any
finite subset $\{x_{0}, \ldots, x_{n-1}\}\subseteq C$ and for any set of non-negative reals $\{\alpha_{0}, \ldots, \alpha_{n-1}\}$
with $\sum_{i<n}\alpha_{i}=1$ , we have
$\sum_{i<n}\alpha_{i}x_{i}\in C$.
PROOF. Fix any $\{x_{0}, \ldots, x_{n-1}\}\subseteq C$ . We first prove the following statement by
induction on $k\leq n$ :
$(*)$ for any non-negative rationals $\{q_{0}, \ldots , q_{k-1}\}$with $\sum_{i<k}q_{i}=1$ ,
$\sum_{i<k}q_{i}x_{i}\in C$.
We here notice that the statement $(*)$ is $\Pi_{1}^{0}$ , and so we can use the $\Pi_{1}^{0}$ induction
(which is equivalent to $\Sigma_{1}^{0}$ induction, see [8, Lemma 1.1]). Assume $(*)$ holds for $k$ .
Let $\{q_{0}, \ldots, q_{k}\}$ be a set of non-negative rationals such that $\sum_{i\leq k}q_{i}=1$ . We may
assume $q_{k}\neq 1$ . For if $q_{k}=1$ then $q_{i}=0$ for $i<k$ , and so $\sum_{i\leq k}q_{i}x_{i}=x_{k}\in C$ . Now
consider the set of rationals $\{\frac{q_{0}}{1-q_{k}}, \frac{q_{1}}{1-q_{k}}, \ldots, \frac{q_{k-1}}{1-q_{k}}\}$ . Then we have
$\frac{q_{0}}{1-q_{k}}+\frac{q_{1}}{1-q_{k}}+\ldots+\frac{q_{k-1}}{1-q_{k}}=\frac{\Sigma_{i\leq k}q_{i}-q_{k}}{1-q_{k}}=1$ .
/0
54
So by the induction hypothesis,
$\sum_{i<k}\frac{q_{i}}{1-q_{k}}x_{i}\in$ C.
Finally, by the convexity of $C$ , we have
$\sum_{i\leq k}q_{k}x_{k}=(1-q_{k})(\sum_{i<k}\frac{q_{i}}{1-q_{k}}x_{i})+q_{k}x_{k}\in C$.
Thus, for any non-negative rationals $\{q_{0}, \ldots , q_{n-1}\}$ with $\Sigma_{i<n}q_{i}=1$ ,
$\sum_{i<n}q_{i}x_{i}\in C$.
Since $C$ is closed, we can easily show that for any non-negative reals $\{\alpha_{0}, \ldots, \alpha_{n-1}\}$
with $\Sigma_{i<n}\alpha_{i}=1$ ,
$\sum_{i<n}\alpha_{i}x_{i}\in C$.
This completes the proof. $\square$
4. Uniform continuity and integration
In this section, we discuss some important properties of uniformly continuous
functions, and then define the Riemann integration for those functions.
We begin with the following definition (cf. Aberth [1], Simpson [14]). Let $f$ be a
continuous function from a closed set $C(\subseteq R^{n})$ to $R^{m}$ . Then $f$ is said to be weakly
uniformly continuous on $C$ if for any $e\in N$ , there exists $d\in N$ such that for all $x$
and $y$ in $C$ , if $||x-y\Vert_{n}<2^{-d}$ then 1 $f(x)-f(y)\Vert_{m}<2^{-e}$ . $f$ is said to be uniformly
continuous on $C$ if there exists a total function $h$ : $Narrow N$ such that for all $e\in N$
and all $x$ and $y$ in $C$ , if $\Vert x-y\Vert_{n}<2^{-h(e)}$ then 1 $f(x)-f(y)\Vert_{m}<2^{-e}$ . Such a function
$h$ is called a modulus of uniform continuity for $f$ .
//
55
4.1 LEMMA$(RCA_{0})$ . Let $C$ be a compact closed set in $R^{n}$ . Then any continuous
function $f$ : $Carrow R^{m}$ is weakly uniformly continuous on $C$ .
PROOF. Let $\Phi=\{(B_{i}, B_{i}’) : i\in N\}$ be a code for a continuous function $f$ :
$C(\subseteq R^{n})arrow R^{m}$ , where $B_{i}$ and $B_{i}’$ are basic open sets in $R^{n}$ and $R^{m}$ , respectively.
Fix any $\epsilon>0$ . Let $\mathcal{B}=$ { $B$ : $(B,$ $B’)\in\Phi$.and $B’=(a,$ $r)$ with $r< \frac{\epsilon}{2}$ }. From the
definition of the domain of a continuous function, it is easy to see that 6 is an open
covering of $C$ . So by the compactness of $C$ , there exists a finite subset of $\mathcal{B}$ which
also covers $C$ . Let $\{(a_{0}, r_{0}), (a_{1}, r_{1}), \ldots, (a_{k-1}, r_{k-1})\}$ be such a finite covering. Now
put $C=\{(a_{i}, r_{i}-2^{-l}) : r_{i}-2^{-l}>0, i<k, l\in N\}$ . It can be shown that $C$ is an
open covering of $C$ , too. Again by the compactness of $C$ , there is an $L\in N$ such
that $C_{L}=\{(a_{i}, r_{i}-2^{-l}) : r_{i}-2^{-l}>0, i<k, l\leq L\}$ also covers $C$ . We finally set
$\delta=2^{-L}$ , and show that for all $x$ and $y$ in $C$ , if $\Vert x-y\Vert_{n}<\delta$ then 1 $f(x)-f(y)\Vert_{m}<\epsilon$ .
Choose any two points $x$ and $y$ from $C$ such that $\Vert x-y\Vert_{n}<\delta$ . Since $C_{L}$ covers $C$ ,
there exists a basic open set $(a_{i}, r_{i}-2^{-l})$ in $C_{L}$ which the point $x$ belongs to. Then
both $x$ and $y$ belong to the basic open set $(a_{i}, r_{i})$ in $B$ , since $\Vert x-y\Vert_{n}<2^{-L}\leq 2^{-l}$ .
Hence, by the definition of $B$ , both $f(x)$ and $f(y)$ belong to a basic open set $(a, r)$
with $r< \frac{\epsilon}{2}$ that is, $\Vert f(x)-f(y)\Vert_{m}<\epsilon$ . This completes the proof. $\square$
4.2 LEMMA$(WKL_{0})$ . Let $C$ be an n-dimensional rectangle { $(x_{0}, \ldots, x_{n-1})\in R^{n}$ :
$a_{i}\leq x_{i}\leq b_{i}\}$ with $a_{i},$ $b_{i}\in R$ . Then any continuous function $f$ : $Carrow R^{m}$ is uniformly
continuous on $C$ .
PROOF. Let $f$ be a continuous function from the rectangle $C$ to $R^{m}$ . Within
$WKL_{0}$ , a rectangle $C$ is compact by Lemma 3.1. So we know from Lemma 4.1, that
for each $e$ , there exists $d$ such that $\Vert x-y\Vert_{n}<2^{-d}\Rightarrow\Vert f(x)-f(y)\Vert_{m}<2^{-e}$ . Look back
at the proof of Lemma 4.1. In the proof, the compactness of $C$ is used twice, first for
$/Z$
56
the covering $\mathcal{B}$ and second for $C$ . If $C$ is a rectangle $[a_{0}, b_{0}]\cross[a_{1}, b_{1}]\cross\ldots\cross[a_{n-1}, b_{n-1}]$
with $a_{i},$ $b_{i}\in Q$ , we can easily decide whether a given finite subset of $\mathcal{B}$ (and $C$ ) covers
$C$ or not, and thus obtained $d$( $=L$ in the proof of Lemma 4.1) from $e$ in a recursive
way. We now want to reduce the general case $(a_{i}, b_{i}\in R)$ to this special case.
Suppose for each $i<n,$ $a_{i}=\{a_{ij} : j\in N\},$ $b_{i}=\{b_{ij} : j\in N\}$ with $a_{ij},$ $b_{ij}\in Q$ . For
each $j\in N$ , let $C_{j}$ be the rectangle $\{(x_{0}, \ldots, x_{n-1})\in R^{n} : a_{ij}-2^{-j}\leq x_{i}\leq b_{ij}+2^{-j}\}$.
Then it is easy to see that $C= \bigcap_{j\in N}C_{j}$ . Define $\mathcal{B}$ as in the proof of Lemma 4.1. Since
$\mathcal{B}$ is an open covering of $C$ and $C$ is compact, there exists $j_{0}\in N$ such that $\mathcal{B}$ covers
$\bigcap_{j<j_{0}}C_{j}$ . So we can effectively find a finite subset $\{(a_{0}, r_{0}), (a_{1}, r_{1}), \ldots , (a_{k-1}, r_{k-1})\}\subseteq$
$\mathcal{B}$ which covers $\bigcap_{j<k}C_{j}$ . Define $C$ as before. Now $C$ is an open covering of $\bigcap_{j<k}C_{j}$ .
We can then find a finite subcovering $C_{L}$ in an effective way. Thus, in $WKL_{0}$ , there
exists a total function $h:Narrow N$ such that for all $e\in N$ and for all $x,$ $y\in C$ ,
$\Vert x-y\Vert_{n}<2^{-h(e)}\Rightarrow\Vert f(x)-f(y)\Vert_{m}<2^{-e}$ . $\square$
For the reversal of the above lemma, see Aberth [1, Theorem 7.3] and Simpson
[13]. They indeed show that $WKL_{0}$ and the statment that any continuous function
on $[0,1]$ is uniformly continuous are equivalent over $RCA_{0}$ .
The next lemma shows that a uniformly continuous function from a countably
represented closed set $\hat{A}(\subseteq R^{n})$ to $R^{m}$ can be uniquely encoded by its restriction to
$A$ . Since a function from $A$ to $R^{m}$ is just a point in $(R^{m})^{N}\approx R^{N}$ , this lemma is very
useful to deal with certain function spaces (cf. the discussion before Theorem 6.2).
4.3 LEMMA$(RCA_{0})$ . Let $C$ be a closed set in $R^{n}$ . Suppose that $C$ is countably
represented by $A\subseteq Q^{n}$ . Let $f$ : $Aarrow R^{m}$ be a uniformly continuous function with a
modulus function $h$ (i.e., for all $a$ and $b$ in $A,$ $\Vert a-b\Vert_{n}<2^{-h(e)}\Rightarrow\Vert f(a)-f(b)\Vert_{m}<$
$2^{-e})$ . Then there exists (a code for) a unique continuous function $\hat{f}$ : $\hat{A}arrow R^{m}$ such
/3
’ $*$ 81$j$ $r$
that $f(a)=f(a)$ for all $a$ in $A$ .
PROOF. Let $f$ : $Aarrow R^{m}$ be a uniformly continuous function with modulus $h$ . Let
$\Phi(\subseteq Q^{n}\cross Q^{+}\cross Q^{m}\cross Q^{+})$ be a code for a continuous function $\hat{f}$ : $\hat{A}arrow R^{m}$ defined
by
$(a, r, b, s)\in\Phirightarrow\exists e$ ( $r<2^{-h(e)}$ and 1 $f(a)-b\Vert_{m}<s-2^{-e}$).
Note that the above definition is $\Sigma_{1}^{0}$ , and so $\Phi$ can be seen as a recursive enumeration
of its members. It is easy to see that $\Phi$ satisfies all the conditions to be a continuous
function code. It is also clear that $f(a)=f(a)$ for all $a$ in A. $\square$
We next discuss the integration of uniformly continuous function on a closed in-
terval $[a, b]$ with $a,$ $b\in R$ . $For$ simplicity, we do not deal with multi-variable functions
here. Recall that a continuous function on { $a,$$b$ ] is always uniformly continuous in
$WKL_{0}$ , but not always in $RCA_{0}$ .
Let $f$ : $[a, b]arrow R$ be uniformly continuous with a modulus function $h$ . We define
a function $S:Narrow R$ by
$S(m)= \frac{b-a}{2^{m}}\sum_{k=0}^{2^{m}-1}f(a+k\frac{b-a}{2^{m}})$ .
Then the (Riemann) integration of $f$ on $[a, b]$ , denoted $\int_{a}^{b}f(x)dx$ , is defined to be
$\lim_{narrow\infty}S(h(n))$ . To see the existence of such a limit, we first show the following
lemma.
4.4 LEMMA$(RCA_{0})$ . If a sequence $\{x_{n}\}\in R$ satisfies $\exists n_{0}\forall n\forall i(|x_{n}-x_{n+i}|\leq$
$2^{n0-n})$ , then there exists $x\in R$ such that $\forall m\exists M\forall n\geq M(|x-x_{n}|\leq 2^{-m})$ . (cf.
Brown and Simpson [3, Lemma 4.1]).
PROOF. For $n\in N$ , let $x_{n}=\langle q_{n,k}$ : $k\in N$ } with $q_{n,k}\in Q$ . Let $x=\langle q_{n+no+2,n+n_{0}+2}$ :
$n\in N)$ . Then it is easy to see that $x\in R$ and $\forall n\geq m+n_{0}+1(|x-x_{n}|\leq 2^{-m})$ . $\square$
$/+$
58
Choose $m_{0}\in N$ such that $|b-a|\leq 2^{m_{0}}$ . Then by the uniform continuity of $f$ , for
any $x \in[a+k\frac{b-a}{2^{h\langle n)+m_{0}}},$ $a+(k+1) \frac{b-a}{2^{h(n)+mo}})$ , we have
$|f(x)-f(a+k \frac{b-a}{2^{h(n)+m_{0}}})|<2^{-n}$ .
Hence, for all $i\in N$ ,
$|S(h(n+i)-m_{0})-S(h(n)+m_{0})|< \frac{|b-a|}{2^{h(n+i)+m_{0}}}(2^{h\langle n+i)+m_{0}}\cdot 2^{-n})$
$=|b-a|\cdot 2^{-n}\leq 2^{m0-n}$ .
Therefore, by the above lemma, $\lim_{narrow\infty}S(h(n))=\lim_{narrow\infty}S(h(n)+m_{0})$ exists.
5. Brouwer’s fixed point theorem
In this section, we investigate what set existence axioms are needed to prove some
variants of Brouwer’s fixed point theorem. We begin with the following theorem.
5.1 BROUWER’S THEOREM I $(a)(RCA_{0})$ . For any continuous function $f$ : $[0,1]arrow$
$[0,1]$ , there is a point $x\in[0,1]$ such that $f(x)=x$ .
$(b)(WKL_{0})$ . For any continuous function $f$ : $[0,1]^{n}arrow[0,1]^{n}$ , there is a point
$x\in\lfloor 0,1]^{n}$ such that $f(x)=x(n\geq 2)$ .
PROOF. (a). We imitate Simpson’s proof for the intermediate value theorem [15].
Suppose that for all rational $q\in[0,1],$ $f(q)\neq q$ . With the $\triangle_{1}^{0}$ comprehension, we
define a nested sequence of rational intervals follows:
$[a_{0}, b_{0}]=[0,1]$
$[a_{n+1}, b_{n+1}]=\{\begin{array}{l}[(a_{n}+b_{n})/2,b_{n}]iff((a_{n}+b_{n})/2)>(a_{n}+b_{n})/2[a_{n},(a_{n}+b_{n})/2]iff((a_{n}+b_{n})/2)<(a_{n}+b_{n})/2\end{array}$
$/\cdot\sigma$
59
By nested interval convergence (see Lemma 2.2 in Simpson [15]), there exists a real
$x$ such that $x= \lim a_{n}=\lim b_{n}$ . This $x$ is a fixed point for $f$ .
(b). Among several known proofs of this theorem, one given by D. Gale[9] seems
to be most easily carried out within $RCA_{0}$ . His proof nostly consists of manipulations
of finite objects (HEX), which do not use any set existence axiom. The only infinitary
argument or fact you need is that every continuous function on $[0,1]^{n}$ is uniformly
continuous, which is proved in our Lemma 4.2. For details, see Gale[9]. $\square$
We remark that part (b) is not provable over $RCA_{0}$ . In fact, we have
5.2 THEOREM $(RCA_{0})$ . The following are pairwise equivalent:
(i) $WKL_{0}$ ,
(ii) for any continuous function $f$ : $[0,1]^{2}arrow[0,1]^{2}$ , there is a point $x\in[0,1]^{2}$ such
that $f(x)=x$ .
PROOF. Since $(i)arrow(ii)$ is already proved by Theorem 5.1.(b), we only show
(ii) $arrow(i)$ . Our argument is essentially due to V. P. Orekov (cf. chapter IV of
Beeson[2]).
By way of contradiction, deny (i). Then $[0,1]$ is not compact by Lemma 3.1.
Let $\{I_{i} : i\in N\}$ be a sequence of rational open intervals which covers $[0,1]$ but has
no finite subcover. From $\langle I_{i}\rangle$ , we can easily construct a sequence of rational closed
intervals ( $J_{i}$ : $i\in N\rangle$ such that $\phi\neq J_{i}\subseteq[0,1],$ \langle $J_{i}$ } covers $[0,1]$ , and any two distinct
$J_{i}’s$ are disjoint or have only an endpoint in common. We may assume that $J_{0}$ has $0$
as its left endpoint and $J_{1}$ has 1 as its right endpoint. Define $A_{k}$ to be the union of
all $J_{i}\cross J_{k}$ and $J_{k}\cross J_{i}$ for $i\leq k$ .
From now, we construct a retraction $f$ of $[0,1]^{2}$ onto the four side of the square
$[0,1]^{2}$ . If such an $f$ is constructed, (ii) does not hold. For if $r$ is the rotation of $90^{O}$
/6
60
about the point $($ }, $\frac{1}{2}),$ $rof$ is a continuous function from $[0,1]^{2}$ into itself which has
no fixed point.
We define a retraction $f$ in stages. Suppose $f$ has defined on all $A_{i}$ for $i<k$ . We
want to define $f$ for $A_{k}$ , compatibly with the value already assigned. Decompose $A_{k}$
into its connected components $P_{1},$ $\ldots P_{l}$ . We can easily observe that each $P_{i}$ has at
least one free side on which $P_{i}$ does not adjoin $\bigcup_{i<k}A_{i}$ or any side of $[0,1]^{2}$ . Now
$f$ can be extended to $P_{i}$ by combining a retraction of $P_{i}$ onto the sides on which
the values of $f$ are already determined (or onto any one side if no such side). Since
$\bigcup_{k\in N}A_{k}=[0,1]^{2}$ , this procedure clearly defines a retraction of $[0,1]^{2}$ onto its sides.
More formally, we have to construct a code for this retraction. This is done by
enumerating the pairs of nonempty basic open sets $(B_{1}, B_{2})$ such that $B_{1} \subseteq\bigcup_{i<k}A_{i}$
for some $k$ and $f(B_{1})\subseteq\overline{B_{2}}$. Since this construction is just a routine, we omit the
details. $\square$
We now generalize Theorem 5.1 as follows:
5.3 BROUWER’S THEOREM $II(WKL_{0})$ . Let $\{a_{0}, \ldots , a_{k}\}$ be a finite subset of $Q^{n}$ .
Let $C$ be the closed set { $\sum_{i\leq k}\alpha_{i}a_{i}$ : $\alpha_{i}\in R,$ $\alpha_{i}\geq 0$ for $i\leq k$ , and $\Sigma_{i\leq k}\alpha_{i}=1$ }. Then
for any continuous function $f$ : $Carrow C$ , there is a point $x\in C$ such that $f(x)=x$ .
PROOF. As in the standard proof (see [17]), we will show that $C$ is a retract of
a sufficient large n-dimensional rectangle (in a certain coordinate system). Changing
coordinate systems is not essential, but makes it much easier to construct such a
retraction.
We first find a basis for the space $L$ spanned by $\{a_{1}-a_{0}, a_{2}-a_{0}, \ldots , a_{k}-a_{0}\}$ . This
can be done by simple calculations of rational matrices (e.g., Gaussian elimination).
Notice that the assumption $\{a_{0}, a_{1}, \ldots, a_{k}\}\subseteq Q^{n}$ (rather $than\subseteq R^{n}$ ) is necessary to
/7
61
determine whether some entries of the matrices in the computations are zero or not.
Let $\mathcal{B}$ be a base for $R^{n}$ including the base for the subspace $L$ spanned by $\{a_{1}-$
$a_{0},$ $a_{2}-a_{0},$ $\ldots,$ $a_{k}-a_{0}$ }. Using the coordinate system relative to the base $B$ , the points
$a_{0},$ $a_{1},$ $\ldots,$$a_{k}$ can be expressed as n-tuples in $R^{l}\cross\{0\}^{n-l}$ , where $I$ is the dimension of
the subspace $L$ . So there exists $d\in R$ such that the convex hull $C$ of $\{a_{0}, a_{1}, \ldots, a_{k}\}$
is included in $[-d, d]^{l}\cross\{0\}^{n-l}\subseteq[-d, d]^{n}$ with respect to the new coordinate system.
There is an obvious retraction from $[-d, d]^{n}$ onto $[-d, d]^{l}\cross\{0\}^{n-l}$ . So if we can show
that $C$ is a retract of $[-d, d]^{l}\cross\{0\}^{n-l}$ , then we may conclude that $C$ has the fixed
point property (i.e., any continuous function from $C$ to itself has a fixed point) since
we already know from Theorem 5.1 that $[-d, d]^{n}$ has the fixed point property. Remark
that if $C$ has the fixed point property in the new coordinate system then it also has
in the standard coordinate system, since a continuous function code in the standard
coordinate system can be easily translated in the new coordinate system (and vice
versa).
From now on, we regard $C$ as a subset of $[-d, d]^{l}$ by ignoring the zeros in the $(l+1)-$
th to n-th coordinates. To clarify the shape of $C$ , we remove all the superfluous points
from $\{a_{0}, a_{1}, \ldots, a_{k}\}$ and construct the smallest subset $S\subseteq\{a_{0}, a_{1}, \ldots, a_{k}\}$ such that
the convex hull of $S$ is still $C$ . Note that $a_{i_{0}}$ is superfluous if there are $l+1$ points
$a_{i_{1}},$ $a_{i_{2}}$ , , .. , $a_{i_{1}+1}$ in $\{a_{0}, a_{1}, \ldots, a_{k}\}$ such that $a_{i_{0}}\neq a_{i_{j}}$ for all $j\neq 0$ , and such that $a_{i_{0}}$
is involved in the convex hull of $\{a_{i_{1}}, a_{i_{2}}, \ldots, a_{i_{l}+1}\}$ .
We may assume that $\{a_{0}, \ldots, a_{k}\}$ has no superfluous points. Let $\overline{a}$ be the center
of $C$ , i.e., $\overline{a}=(\Sigma_{i\leq k}a_{i})/(k+1)$ . We construct a retraction $\hat{g}$ : $[-d, d]^{l}arrow C$ as
follows. For $b\in[-d, d]^{l}\cap Q^{l}$ , if $b\in C$ then we put $g(b)=b$ , and if $b\not\in C$ then we
put $g(b)=$ the point at which the line segment connecting $b$ and $\overline{a}$ intersects a face
1 $8’$
62
of $C$ . Note that such an intersection can be obtained by solving a system of linear
equations with rational coefficients. The function $g$ thus defined on $[-d, d]^{l}\cap Q^{l}$ can
be uniquely extended to a continuous function $\hat{g}$ on $[-d, d]^{l}$ . In fact, a code $\Phi$ for the
continuous function $\hat{g}$ is defined by
$(B, B’)\in\Phi$ $\Leftrightarrow$ $\exists b_{0},$ $b_{1},$$\ldots,$
$b_{l}\in[-d, d]^{l}\cap Q^{l}$
$B$ is included in the convex hull of $\{b_{0}, \ldots , b_{l}\}$
and $\{g(b_{0}), \ldots, g(b_{l})\}$ is included $inB’$ .
Then we can easily see that $\Phi$ is indeed a code for the desired retraction. This
completes the proof. $\square$
The above theorem can be further generalized to Theorem 5.4 and Theorem 6.1.
Since the next theorem can be proved in the same way as Theorem 6.1, we just state
it without proof.
5.4 BROUWER’S THEOREM $III(WKL_{0})$ . Let $C$ be a nonempty compact convex
closed set in $R^{n}$ , which is also assumed to be countably represented by $A\subseteq Q^{n}$ . Then
for any continuous function $f$ : $Carrow C$ has a fixed point.
6. Tychonoff-Schauder theorem for $R^{N}$ and its applications to
ordinary differential equations
In this section, we extend Brouwer’s theorem to its infinite dimensional analogue
(Tychonoff-Schauder theorem for $R^{N}$ ), from which we prove the Cauchy-Peano the-
orem for ordinary differential equations.
6.1 TYCHONOFF-SCHAUDER THEOREM$(WKL_{0})$ . Let $C$ be a nonempty compact
convex closed set in $R^{N}$ , which is also assumed to be countably represented by $A\subseteq$
/7
63
$Q^{N}$ . Then for any continuous function $f$ : $Carrow C$ has a fixed point.
PROOF. By way of contradiction, we assume that for all $x\in C,$ $f(x)\neq x$ . Let $\Phi$ be
a code for $f$ . Recall that $(B, B’)\in\Phi$ means that $f(B)\subseteq\overline{B’}$. Let $B=\{B$ : there is $B’$
such that $(B, B’)\in\Phi$ and $B\cap B’=\emptyset$ }. It is easy to see that $\mathcal{B}$ covers $C$ , since $f(x)\neq$
$x$ for all $x\in C$ . By the compactness of $C,$ $l3$ has a finite subcover $\{B_{0}, B_{1}, \ldots, B_{k}\}$ .
So there exists $n\geq 1$ and $\epsilon>0$ such that for all $x\in C$ , I $f(x)-x\Vert_{n}>\epsilon$ . We may
assume $\epsilon$ is a rational number. Fix $n$ and $\epsilon$ . For contradiction, we will show that
there exists $x\in C$ such that 1 $f(x)-x\Vert_{n}\leq\epsilon$ . We define, for each $a\in A$ ,
$Ta=\{B$ : there exists $B’=(b, s)$ with $\dim(b)\geq n$
such that $(B, B’)\in\Phi$ and $\Vert b-a\Vert_{n}+s\leq\epsilon$}.
$Ta$ is a $\Sigma_{1}^{0}$ predicate and hence $Ta$ can be viewed as a recursive enumeration of its
numbers (cf. Lemma 2.1 of Simpson [14]). Clearly, $\cup\{Ta : a\in A\}$ covers $C$ . So by
compactness, it has a finite subcover $\{B_{ij} : i\leq k,j\leq l\}$ such that $B_{ij}\in Ta_{i}$ , where
$a_{i}$ is the $(i+1)$ -th element of $A$ . From this subcover, we will construct a continuous
function $g$ on a finite dimensional set, which has a fixed point by Brouwer’s theorem.
Then from this fixed point, we will make $x\in C$ such that 1 $f(x)-x\Vert_{n}\leq\epsilon$ .
Suppose that $B_{ij}=(b_{ij}, r_{ij})$ and $\dim(b_{ij})=m_{ij}$ for $i\leq k$ and $j\leq l$ . Let
$m= \max$ [ $\{m_{ij}$ : $i\leq k$ and $j\leq l\}\cup\{n\}$ ]. For $x\in R^{N},$ $x[m]$ denotes the initial
segment of $x$ with length $m$ . Let $D$ be the convex hull of $\{a_{i}[m] : i\leq k\}$ . We will
construct a continuous function $g$ on $D$ .
For each $i\leq k$ , we first define a continuous function $d_{i}$ : $Darrow R$ by
$d_{i}(x)= \max[\{r_{ij}-\Vert x-b_{ij}\Vert_{m_{ij}} : j\leq l\}\cup\{0\}]$ .
Note that $d_{i}(x)>0$ iff any extension of $x$ to $C$ belongs to $B_{ij}$ for some $j\leq l$ . It is not
2 $0$
64
difficult (but somewhat messy) to construct a code for the continuous function $d_{i}$ . So
we leave this construction to the reader. It is also easy to see that $\sum_{i\leq k}d_{j}(x)>0$ for
each $x\in D$ . We set
$\beta_{i}(x)=\frac{d_{i}(x)}{\Sigma_{j\leq k}d_{j}(x)}$ .
Then for all $x\in D$ , we have
$\sum_{i\leq k}\beta_{i}(x)=1$.
We finally define a continuous function $g:Darrow D$ by
$g(x)= \sum_{i\leq k}\beta_{i}(x)a_{i}[m]$.
We now apply Brouwer’s theorem II to the function $g:Darrow D$ . Let $x$ be any fixed
point of $g$ , and $\overline{x}$ be a point in $C$ such that $\overline{x}[m]=x$ . Let $I=\{i\leq n : \beta_{i}(x)>0\}$ .
Such an $I$ exists by bounded $\Sigma_{1}^{0}$ separation (cf. Lemma 1.6 of [8]). We here notice
that
$||f(\overline{x})-a_{i}\Vert_{n}\leq\epsilon$ for all $i\in I$ ,
since if $\beta_{i}(x)>0,\overline{x}$ belongs to $B_{ij}$ for some $j$ and hence $\overline{x}$ belongs to $Ta_{i}$ .
Since $\sum_{i\in I}\beta_{i}(x)=1$ , we have
$|If(\overline{x})-\overline{x}\Vert_{n}$ $=$ I $\sum_{i\in I}\beta_{i}(x)f(\overline{x})-\sum_{i\in I}\beta_{i}(x)a_{i}\Vert_{n}$
$\leq$
$\sum_{i\in I}\beta_{i}(x)\Vert f(\overline{x})-a_{i}\Vert_{n}$
$\leq$ $\epsilon$ .
This completes the proof. $\square$
Remark: The above theorem is indeed provable within $RCA_{0}$ . For if the compact
set $C$ includes a line segment, the line segment is also compact, which implies $WKL_{0}$
by Lemma 3.1, and otherwise the theorem is trivial since $C$ consists of a single point.
2/
65
As an application of the above fixed point theorem, we prove, within $WKL_{0}$ ,
the Cauchy-Peano theorem for ordinary differential equations. In [14], Simpson has
already proved the Cauchy-Peano theorem in $WKL_{0}$ by eliminating the use of the
Ascoli lemma from Peano’s original proof. A key idea in his proof is that a solution
of the initial value problem can be found in a set of equicontinuous (or Lipchitzian)
functions, which can be encoded by points in $R^{N}$ (see Lemma 4.3 in this paper). In
order to use the fixed point theorem, we need a precise description of the set in $R^{N}$
corresponding to the set of equicontinuous functions, while Simpson only uses the fact
that the equicontinuous functions can be embedded into the compact set $[$ -1, $1]^{N}$ .
We here emphasize that our aim is not merely to prove the Cauchy-Peano theorem
but also to develop a part of functional analysis related to fixed point theorems within
second-order arithmetic.
6.2 $CAUCHY- PEANOTHEOREM(WKL_{0})$ . Let $f(x, y)$ be a continuous function
from the rectangle $D=\{(x, y)\in R^{2} : |x|\leq a, |y|\leq b\}$ to $R$ . Then the initial value
problem
$\frac{dy}{dx}=f(x, y)$ , $y(0)=0$
has at least one solution $y=\phi(x)$ on the interval $[-\alpha, \alpha]$ , where $\alpha=\min(a, b/M)$
and $M= \max\{|f(x, y)| : (x, y)\in D\}$ . (Note: It is provable in $WKL_{0}$ that $f$ has a
maximum on D. see [13].)
PROOF. Suppose that $\{q_{i}\}_{i\in N}$ enumerates the rationals in $[-\alpha, \alpha]$ so that $q_{0}=0$
and $q_{i}\neq q_{j}(i\neq j)$ . We define a closed set $C$ in $R^{N}$ by
\langle $u_{i}$ } $\in C\Leftrightarrow u_{0}=0$ and $|u_{i}-u_{j}|\leq M|q_{i}-q_{j}|$ for all $i,j$ .
By Lemma 4.3, we may think that $u=\{u_{i}\rangle$ $\in C$ encodes a continuous function
$\tilde{u}$ : $[-\alpha, \alpha]arrow R$ such that $\tilde{u}(q_{i})=u_{i}$ for all $i$ . By identifying $u$ with $\tilde{u},$ $C$ can be
22
66
regarded as the set of continuous functions $g$ : $[-\alpha, \alpha]arrow R$ such that $g(0)=0$ and
$|g(x)-g(y)|\leq M|x-y|$ for all $x,$ $y\in[-\alpha, \alpha]$ . It is easy to see in $WKL_{0}$ that $C$ is
compact and convex. To apply Theorem 6.1 to this $C$ , we also need to show that $C$
is countably represented by some $A\subseteq Q^{N}$ .
Let $P=$ {{ $p_{0},$ $\ldots$ , $p_{n}\rangle\in Q^{<N}$ : $p_{0}=0$ and $|p_{i}-p_{j}|<M|q_{i}-q_{j}|$ if $i\neq j$ }. For
$p=$ \langle $p_{0},p_{1},$ $\ldots$ , $p_{n}\}\in P$ , we define a standard extension $\overline{p}=\{\overline{p}0f\overline{l},$ $\ldots,\overline{p}_{n},\overline{p}_{n+1},$ $\ldots\rangle$
of $p$ into $Q^{N}$ as follows: for each $k\geq 0$ ,
$\overline{p}_{k}=\{\begin{array}{l}p_{i}ifq_{k}\in[q_{i},\alpha],i\leq nand\neg\exists l\leq n(q_{l}\in(q_{i},\alpha])\frac{p_{J}-pi}{q_{j}-q}(q_{k}-q_{i})+p_{i}ifq_{k}\in[q_{i},q_{j}),i,j\leq nand\neg\exists l\leq n(q_{l}\in(q_{i},q_{j}))pjifq_{k}\in[-\alpha,q_{j}),j\leq nand\neg\exists l\leq n(q_{l}\in[-\alpha,q_{j}))\end{array}$
We put $A=\{\overline{p}\in Q^{N} : p\in P\}$ . Then $A$ can be seen as the set of piecewise
linear functions on $[-\alpha, \alpha]$ . We will show that $C$ is countably represented by $A$ .
Choose any $u=\{u_{i}\rangle$ $\in C$ . We want to find a sequence $\{\overline{p}^{k} : k\in N\}$ from $A$
such that $\Vert u-\overline{p}^{k}\Vert_{k+1}\leq 2^{-k}$ for each $k$ . Fix $k\in N$ . We construct a sequence
$p^{k}=\{p_{0},$ $p_{1},$ $\ldots,$$p_{k}\rangle$ in $P$ such that $|u_{i}-p_{i}|\leq 2^{-k}$ for each $i\leq k$ , and then extend it
to $\overline{p}^{k}\in A$ . Let $n\in N$ be large enough (strictly, $n\geq 2k+2$ and $\frac{2^{k+1}}{2^{n}}\leq M|q_{i}-q_{j}|$ for all
$i,j\leq k(i\neq j))$ . For each $i\leq k$ , let $r_{i}$ be a rational number such that $|u_{i}-r_{i}|<2^{-n}$ .
Suppose $|r_{i_{0}}|\leq|r_{i_{1}}|\leq\ldots\leq|r_{i_{k}}$ I with $\{i_{0}, i_{1}, \ldots, i_{k}\}=\{0,1, \ldots, k\}$ . We now define
$\{p_{0},p_{1}, \ldots, p_{k}\}$ as follows: for $j\leq k$ ,
$p_{i_{j}}=\{\begin{array}{l}r_{i_{j}}-\frac{2^{j+1}}{2^{n}}ifr_{i_{j}}\geq\frac{2^{j+1}}{2^{n}}r_{i_{j}}+\frac{2^{j+1}}{2^{n}}ifr_{i_{j}}\leq-\frac{2^{j+1}}{2^{n}}0otherwise\end{array}$
Then it is easy to see that $|u_{i}-p_{i}|\leq 2^{-k}$ for all $i\leq k$ . To show $\{p_{0}, \ldots, p_{k}\}\in P$ , we
23
67
compute: for $j<l\leq k$ ,
$|p_{i_{j}}-p_{i_{l}}|$ $\leq$$\max\{|r_{i_{j}}-r_{i_{l}}|-\frac{2^{l+1}-2^{j+1}}{2^{n}}, \frac{2^{l+1}-2^{j+1}}{2^{n}}\}$
$<$ $\max\{|u_{i_{j}}-u_{i_{l}}|+\frac{1}{2^{n}}+\frac{1}{2^{n}}-\frac{2^{2}-2^{1}}{2^{n}}, \frac{2^{k+1}}{2^{n}}\}$
$\leq$ $M|q_{i_{j}}-q_{i_{\iota}}|$ .
It is also obvious that $p_{0}=0$ . Henc$e\{p_{0}, \ldots , p_{k}\}$ is in $P$ , and so it can be extended
to $\overline{p}^{k}$ in $A$ .
We next define a continuous function $F:Carrow C$ as follows: for $u\in C$ ,
$F(u)= \langle\int_{0}^{qi}f(t,\tilde{u}(t))dt:i\in N\}$ ,
where $\tilde{u}$ is a continuous function encoded by $u$ . It is then obvious that $F(u)\in C$ for
$u\in C$ , since $F\overline{(u}$ )(0) $=0$ and for all $x,$ $y$ in $[-\alpha, \alpha]$ ,
$|F \overline{(u})(x)-F\overline{(u})(y)|=|\int_{x}^{y}f(t,\tilde{u}(t))dt|\leq M|x-y|$.
Formally, a code $\Phi$ for $F$ is given by
$(a, r, b, s)\in\Phi\Leftrightarrow a,$ $b\in P$ and $r,$ $s\in Q^{+}$ ,
and if $n=\dim(a)$ and $-\alpha\leq q_{i_{0}}<q_{i_{1}}\ldots<q_{i_{n-1}}\leq\alpha$
with $\{i_{0}, \ldots, i_{n-1}\}=\{0, \ldots, n-1\}$ ,
then there exists $e\in N$ such that
(i) $\alpha-q_{i_{n-1}},$ $q_{i_{j}}-q_{i_{j-1}}(1\leq j\leq n-1),$ $q_{i_{0}}+\alpha$ are all $< \frac{1}{M\cdot 2^{h(e)+2}}$
(ii) $r<2^{-h(e)-2}$
(iii) 1 $F(\overline{a})-b\Vert_{\dim(b)}<s-\alpha\cdot 2^{-e}$ ,
where $h$ is a modulus of uniform continuity for $f$ . Conditions (i) and (ii) together
implies that for any point $u=\{u_{i}\rangle$ $\in B_{r}(a)\cap C,$ $|u_{i}-\overline{a}_{i}|<2^{-h(e)}$ for all $i$ , where
$2+$
68
$\overline{a}=\langle\overline{a}_{i}$ } is the standard extension of $a$ . Then $|f(t,\tilde{u}(t))-f(t,\simeq a(t))|\leq 2^{-e}$ for all
$t\in[-\alpha, \alpha]$ , and so $|F\overline{(u}$ ) $(x)-F\overline{(\overline{a}})(x)|\leq|x|\cdot 2^{-e}\leq\alpha\cdot 2^{-e}$ . Therefore, $F(u)\in B_{s}(b)$
by (iii), which means that $\Phi$ is a code for $F$ . We leave the details to the reader. At
last, the fixed point of $F$ clearly gives a solution of the initial value problem. $\square$
7. Markov-Kakutani fixed point theorem and Hahn-Banach
theorem
The Markov-Kakutani fixed point theorem asserts the existence of a common fixed
point for certain families of affine mappings. A continuous function $T$ from a convex
closed set $C$ to itself is said to be affine if $T(qx+(1-q)y)=qT(x)+(1-q)T(y)$
whenever $q\in[0,1]\cap Q$ and $x,$ $y\in C$ . This theorem has numerous applications [4],
[17]. Among others, Kakutani [11] has proved the Hahn-Banach theorem from this
type of fixed point theorem (see also [10], [18]). In this section, we prove the Markov-
Kakutani theorem for $R^{N}$ within $RCA_{0}$ , and use it to prove the Hahn-Banach theorem
for separable Banach spaces within $WKL_{0}$ .
7.1 $MARKOV- KAKUTANITHEOREM(RCA_{0})$ . Let $C$ be a nonempty compact con-
vex closed set in $R^{N}$ . Let { $T_{n}\rangle$ be a sequence of continuous functions from $C$ to $C$
such that
(i) for each $n,$ $T_{n}$ is affine on $C$ ,
(ii) for each $m$ and $n,$ $T_{n}oT_{m}(x)=T_{m}oT_{n}(x)(x\in C)$ .
Then there exists $x\in C$ such that $T_{n}(x)=x$ for all $n$ .
PROOF. Let $C$ and { $T_{n}\rangle$ be as in the above statement. For each $n\in N$ , let
$K_{n}=\{x\in R^{N} : T_{n}x=x\}\cap C$ . Formally, we express the complement of $K_{n}$ in $R^{N}$
$2S$
69
as the union of the open set { $B$ : $(B,$ $B’)\in\Phi_{n}$ and $B\cap B’=\emptyset$ } and the complement
of $C$ , where $\Phi_{n}$ is a code for $T_{n}$ . Then $K_{n}$ is a closed set in $R^{N}$ . Our goal is to show
$that\cap\{K_{n} ; n\in N\}$ is nonempty.
Let $K_{n,i}=\{x\in R^{N} : \Vert T_{n}x-x\Vert_{i+1}\leq 2^{-i}\}\cap C$ . We can easily see that $K_{n,i}$ is well
defined as a closed set in $R^{N}$ . Since $K_{n}=\cap\{K_{n,i} : i\in N\}$ for each $n$ , we want to
show $that\cap\{K_{n,i} ; n\in N, i\in N\}\neq\emptyset$ .
By way of contradiction, we $assume\cap\{K_{n,i} : n\in N, i\in N\}=\emptyset$ . $Then\cup$ { the
complement of $K_{n,i}$ in $R^{N}$ : $n\in N,$ $i\in N$ } $=R^{N}\supset C$ . By the compactness of $C$ , there
exist $k\in N$ and $l\in N$ such $that\cup$ { the complement of $K_{n,i}$ in $R^{N}$ : $n\leq k,$ $i\leq l$ } $\supset C$ .
$Hence\cap\{K_{n,i} : n\leq k, i\leq l\}\cap C=\emptyset$ , and $so\cap\{K_{n,i} : n\leq k, i\leq l\}=\emptyset$ , sinc$e$ each
$K_{n,i}\subseteq C$ .
Let $z$ be any element of $C$ . Choose $p\in N$ such that $\Vert x-y\Vert_{l+1}\leq p\cdot 2^{-l}$ for all
$x,$ $y\in C$ . The existence of such a $p$ is clear from the compactness of $C$ . We define a
point $x$ in $C$ by
$x= \frac{1}{p^{k+1}}\sum_{i_{0}=0}^{p-1}\ldots\sum_{i_{k}=0}^{p-1}T_{0^{0}}^{i}\ldots T_{k}^{i_{k}}z$ ,
where $T_{n}^{i+1}(z)=T_{n}(T_{n}^{i}(z))$ and $T_{n}^{0}(z)=z$ . We will show $x\in\cap\{K_{n,i} : n\leq k, i\leq l\}$
for a contradiction. Fix any $n\leq k$ . Putting
$x_{n}= \frac{1}{p^{k}}\sum_{i_{0}=0i_{n}}^{p-1}\ldots\sum_{=-10i_{n}}^{p-1}\sum_{=,+10}^{p-1}\ldots\sum_{i_{k}=0}^{p-1}T_{0}^{i_{0}}\ldots T_{n-1}^{i_{n-1}}T_{n+1}^{i_{n+1}}\ldots T_{k}^{i_{k}}z$ ,
we have
.
$\Vert T_{n}x-x\Vert_{l+1}$ $=$ $\Vert T_{n}(\frac{1}{p}\sum_{i_{n}=0}^{p-1}T_{n^{n}}^{i}x_{n})-\frac{1}{p}\sum_{i_{n}=0}^{p-1}T_{n^{n}}^{i}x_{n}\Vert_{l+1}$
$=$ $\Vert\frac{1}{p}\sum_{i_{n}=0}^{p-1}T_{n^{n}}^{i+1}x_{n}-\frac{1}{p}\sum_{i_{n}=0}^{\dot{p}-1}T_{n^{n}}^{i}x_{n}\Vert_{l+1}$
$=$ $\frac{1}{p}\Vert T_{n^{p}}x_{n}-x_{n}\Vert_{l+1}$
26
$\leq$ $\frac{1}{p}(p\cdot 2^{-l})=2^{-l}$ .
Remark that we can easily show by $\Pi_{1}^{0}$ induction (cf. the proof of Lemma 3.2) that
if $T$ : $Carrow C$ is affine, then for any finite subset $\{x_{0}, \ldots, x_{n-1}\}\subseteq C$ , and for any
set of non-negative reals $\{\alpha_{0}, \ldots, \alpha_{n-1}\}$ with $\sum_{i<n}\alpha_{i}=1$ , we have $f( \sum_{i<n}\alpha_{i}x_{i})=$
$\sum\alpha_{i}f(x_{i})$ . From the above inequality, $x\in K_{n,l}=\cap\{K_{n,i} : i\leq l\}$ . Since $n$ is any
number $\leq k,$ $x\in\cap\{K_{n,i} ; n\leq k, i\leq l\}$ . This is a contradiction. So we are done. $\square$
To state the Hahn-Banach theorem, we need introduce some basic notions on
Banach spaces. We define a code for a separable Banach space to be a nonempty set
$A\subseteq N$ together with $operations+,$ $-:A\cross Aarrow A,$ $\cdot$ : $Q\cross Aarrow A,$ $\Vert\cdot\Vert$ : $Aarrow[0, \infty$ )
and distinguished element $0\in A$ such that $A,$ $+,$ $-,$ $0$ forms a vector space over $Q$ and
$\Vert$ . I satisfies 1 $qa\Vert=|q|\Vert a\Vert$ and $\Vert a+b\Vert\leq\Vert a\Vert+\Vert b\Vert$ for all $a,$ $b\in A,$ $q\in Q$ . A point of the
separable Banach space $\hat{A}$ is defined to be a sequence $\langle a_{n} : n\in N\rangle$ from $A$ , satisfying
$\forall n\forall i(\Vert a_{n}-a_{n+i}\Vert\leq 2^{-n})$ . Let $\hat{A}$ and $\hat{B}$ be a separable Banach spaces. A continuous
function $F$ : $\hat{A}arrow\hat{B}$ is encoded as a sequence of $(a, r, b, s)\in A\cross Q^{+}\cross B\cross Q^{+}$
satisfying the three conditions analogous to those for a continuous function on $R^{n}$ .
A bounded Zinear operator is a continuous function $F:\hat{A}arrow\hat{B}$ such that
(i) $F(pa+qb)=pF(a)+qF(b)$ for all $a,$ $b\in A$ and $p,$ $q\in Q$ ,
(ii) there exists a $re$al number $\alpha$ such that $\Vert F(x)\Vert\leq\alpha\Vert a\Vert$ for all $a\in A$ .
If $\alpha\in R$ satisfies $\Vert F(a)\Vert\leq\alpha\Vert a\Vert$ for all $a\in A$ , we write $\Vert F\Vert\leq\alpha$ . A bounded linear
operator $F:\hat{A}arrow\hat{B}$ is also called a bounded linear functional if $\hat{B}=R$ . A bounded
linear operator $F:\hat{A}arrow\hat{B}$ can be encoded by the restriction of $F$ to $A$ (see Lemma
5.4 in Brown and Simpson[3] and Lemma 4.3 in this paper). If there is a bounded
27
71
linear operator $\Psi$ from $\hat{A}$ to $\hat{B}$ satisfying $\Vert\Psi(a)\Vert=\Vert a\Vert$ for all $a\in A$ , then $\hat{A}$ is called
a closed separable subspace of $\hat{B}$ , and a point $x$ in $\hat{A}$ is identified with $\Psi(x)$ in $\hat{B}$ .
We now state the Hahn-Banach theorem for separable Banach spaces, and prove
it within $WKL_{0}$ .
7.2 $HAHN- BANACHTHEOREM(WKL_{0})$ . Let $\hat{S}$ be a closed linear subspace of the
separable Banach space $\hat{A}$ . Let $F:\hat{S}arrow R$ be a bounded linear functional such that
$\Vert F\Vert\leq 1$ . Then there exists a bounded linear functional $\tilde{F}$ : $\hat{A}arrow R$ extending $F$ and
such that $\Vert\tilde{F}\Vert\leq 1$ .
PROOF. Let $\hat{S}$ and $\hat{A}$ be separable Banach spaces. Let $\Psi$ : $\hat{S}arrow\hat{A}$ be a bounded
linear operator satisfying $\Vert\Psi(x)\Vert=\Vert x\Vert$ for all $x\in S$ . Let $F:\hat{S}arrow R$ be a bounded
linear functional such that $\Vert F\Vert\leq 1$ . We want to obtain a bounded linear functional
$\tilde{F}$ : $\hat{A}arrow R$ such that $F(x)=\tilde{F}(\Psi(x))$ for all $x\in\hat{S}$ , and such that $\Vert\tilde{F}\Vert\leq 1$ .
Suppos $e$ that $A=\{a_{i}\},$ $S=\{s_{i}\},$ $a_{0}=0$ and $s_{0}=0$ . We define a closed set $C_{0}$ in
$R^{N}$ by
{ $x_{i}\rangle$ $\in C_{0}\Leftrightarrow x_{0}=0$ and $|x_{i}-x_{j}|\leq\Vert a_{i}-a_{j}\Vert$ for all $i,j$ .
We can easily see in $WKL_{0}$ that $C_{0}$ is compact and convex. Let $P=\{\{u_{0},$$\ldots,$
$u_{n}$ ) $\in$
$Q^{<N}$ : $u_{0}=0$ and ( $|u_{i}-u_{j}|<\Vert a_{i}-a_{j}\Vert$ or $u_{i}=u_{j}$ )} and $X=\{\{\overline{u}_{i}\}\in R^{N}$ :
$\{u_{0}, \ldots, u_{n}\}\in P$ and $\overline{u}_{i}=\min\{u_{j}+\Vert a_{i}-a_{j}\Vert : j\leq n\}\}$ . As in the proof of Theorem
6.2, we can easily show that $C_{0}$ is countably represented by $X$ . By (a generalization
of) Lemma 4.3, we can identify $g=\{g_{i}\}\in C_{0}$ with the continuous function $\tilde{g}$ : $\hat{A}arrow R$
such that $\tilde{g}(a_{i})=g_{i}$ . So $C_{0}$ may be regarded as the set of continuous functions
$\tilde{g}$ : $\hat{A}arrow R$ such that $\tilde{g}(a_{0})=0$ and $|\tilde{g}(x)-\tilde{g}(y)|\leq\Vert x-y\Vert$ for all $x,$$y\in\hat{A}$ . A desired
functional $\tilde{F}$ will be found in $C_{0}$ .
,2, 8
72
Our proof goes as follows. We define three closed subsets of $C_{0},$ $C_{1}\supset C_{2}\supset C_{3}$ .
Roughly speaking, $C_{1}$ is the set of continuous functions $\tilde{f}$ in $C_{0}$ extending $F,$ $C_{2}$ the
set of continuous functions $\tilde{f}$ in $C_{1}$ such that $\tilde{f}(x+y)=\tilde{f}(x)+\tilde{f}(y)$ for all $x\in\hat{A}$ and
$y\in\hat{S},$ $C_{3}$ the set of continuous functions $\tilde{f}$ in $C_{2}$ such that $\tilde{f}(x+y)=\tilde{f}(x)+\tilde{f}(y)$ for
all $x\in\hat{A}$ and $y\in\hat{A}$ , and such that $\tilde{f}(\alpha x)=\alpha\tilde{f}(x)$ for all $\alpha\in R$ and $x\in\hat{A}$ . Clearly,
any function in $C_{3}$ is a desired functional $\tilde{F}$ . So what we need is to show that $C_{3}$ is
nonempty. To show the non-emptiness of $C_{2}$ and $C_{3}$ , we will apply Theorem 7.1 to
certain families of continuous functions on $C_{1}$ and $C_{2}$ , respectively.
First let $C_{1}=$ { $f\in C_{0}$ : $\tilde{f}(\Psi(s_{i}))=F(s_{i})$ for all $i$ }. Formally, we should express
the complement of $C_{1}$ as an infinite sequence of (codes for) basic open sets. We
however omit this routine work. To prove that $C_{1}$ is not empty, we set $C_{1,n}=\{f\in$
$C_{0}$ : $\tilde{f}(\Psi(s_{i}))=F(s_{i})$ for all $i\leq n$ } and show $C_{1,n}\neq\emptyset$ for each $n$ . Choose any $n$ . We
defin$e$ a point \langle $x_{k}$ } $\in R^{N}$ by
$x_{k}= \min\{F(s_{i})+\Vert a_{k}-\Psi(s_{i})\Vert : i\leq n\}$.
Then we have
$x_{k}\leq F(s_{0})+\Vert a_{k}-\Psi(s_{0})\Vert=\Vert a_{k}\Vert$ ,
$x_{k} \geq\min\{-\Vert\Psi(s_{i})\Vert+\Vert a_{k}-\Psi(s_{i})\Vert : i\leq n\}\geq-$il $a_{k}\Vert$ .
In particular, $x_{0}=\Vert a_{0}\Vert=0$ . We also have
$x_{k}-x_{l}$ $\leq$ $\min\{F(s_{i})+\Vert a_{k}-a_{l}\Vert+\Vert a_{l}-\Psi(s_{i})\Vert : i\leq n\}-x_{l}$
$=$ $\Vert a_{k}-a_{l}\Vert$ ,
$x_{k}-x_{l}$ $\geq$ $x_{k}- \min\{F(s_{i})+\Vert a_{l}-a_{k}\Vert+\Vert a_{k}-\Psi(s_{i})\Vert : i\leq n\}$
$=$ $-\Vert a_{k}-a_{l}\Vert$ .
2 (\dagger
73
Thus $\langle x_{k}\rangle\in C_{0}$ . Let $\tilde{g}$ : $\hat{A}arrow R$ be the continuous function such that $\tilde{g}(a_{k})=x_{k}$ .
Then for each $j\leq n$ ,
$\tilde{g}(\Psi(s_{j}))\leq F(s_{j})+\Vert\Psi(s_{j})-\Psi(s_{j})\Vert=F(s_{j})$,
$\tilde{g}(\Psi(s_{j}))\geq\min\{F(s_{i})+F(s_{j}-s_{i}):i\leq n\}=F(s_{j})$ .
So $g\in C_{1,n}$ . Since $C_{0}$ is compact and $C_{1}= \bigcap_{n}C_{1,n}$ , it follows that $C_{1}\neq\emptyset$ . Moreover,
it is obvious that $C_{1}$ is compact and convex.
Next let $C_{2}=$ { $f\in C_{1}$ : $\tilde{f}(a_{i}+\Psi(s_{j}))=\tilde{f}(a_{i})+\tilde{f}(\Psi(s_{j}))$ for all $i,j$ }. We want to
show $C_{2}\neq\emptyset$ . Define a family of continuous functions $\{T_{j} : C_{1}arrow C_{1}\}$ by
$(T_{j}f)(a_{i})=\tilde{f}(a_{i}+\Psi(s_{j}))-\tilde{f}(\Psi(s_{j}))$ .
Formally, a code $\Phi$ for { $T_{j}\rangle$ is given by
$(j, u,p, v, q)\in\Phi\Leftrightarrow j\in N,$ $u,$ $v\in P$ and $p,$ $q\in Q^{+}$ ,
and if $n=\dim(u),$ $m=\dim(v)$ and $\Psi(s_{j})=\{a_{j_{k}}$ : $k\in N\rangle$ then
(i) $\forall i<m\exists l<n(a_{l}=a_{i}+a_{j_{m-1}})$,
(ii) $p<2^{-m+1}$ ,
(iii) $\Vert T_{j}\overline{u}-v\Vert_{m}<q-6\cdot 2^{-m+1}$ ,
where $\overline{u}=\{\overline{u}_{i}\rangle$ is defined by $\overline{u}_{i}=\min\{u_{j}+\Vert a_{i}-a_{j}\Vert : j<n\}$ . We omit checking that
$\Phi$ really encodes $\langle T_{j}\rangle$ . We show that the range of $T_{j}$ is included in $C_{1}$ as follows: for
each $f\in C_{1},$ $T_{j}f\in\hat{X}$ , since
$|(T_{j}f)(a_{i})-(T_{j}f)(a_{k})|$ $=$ $|\tilde{f}(a_{i}+\Psi(s_{j}))-\tilde{f}(a_{k}+\Psi(s_{j}))|$
$\leq$ $\Vert a_{i}-a_{k}\Vert$ ,
$30$
74
and for each $f\in C_{1},$ $T_{j}f\in C_{1}$ , since
$(T_{j}f)(\Psi(s_{k}))$ $=$ $\tilde{f}(\Psi(s_{k})+\Psi(s_{j}))-\tilde{f}(\Psi(s_{j}))$
$=$ $F(s_{k}+s_{j})-F(s_{j})$
$=$ $F(s_{k})$ .
It is obvious that each $T_{j}$ is affine and $T_{j}oT_{k}=T_{k}oT_{j}$ . So by Theorem 7.1, there
exists $g\in C_{1}$ such that $\tilde{g}(a_{i})=\tilde{g}(a_{i}+\Psi(s_{j}))-\tilde{g}(\Psi(s_{j}))$ for all $i,j$ . Then $C_{2}\neq\emptyset$ .
Moreover, $C_{2}$ is compact and convex.
Finally, let $C_{3}=$ { $f\in C_{2}$ : $\tilde{f}(a_{i}+a_{j})=\tilde{f}(a_{i})+\tilde{f}(a_{j})$ for all $i,j$ }. Define a family
of continuous functions $\{U_{j} : C_{2}arrow C_{2}\}$ by
$(U_{j}f)(a_{i})=\tilde{f}(a_{i}+a_{j})-\tilde{f}(a_{j})$ .
A code for $\{U_{j}\}$ can be encoded in the same way as for \langle $T_{j}$ }. It is easy to see that for
each $f\in C_{2},$ $U_{j}f\in C_{0}$ . To see $U_{j}f\in C_{1}$ , we have
$(U_{j}f)(\Psi(s_{i}))=\tilde{f}(\Psi(s_{i})+a_{j})-\tilde{f}(a_{j})=\tilde{f}(\Psi(s_{i}))=F(s_{i})$.
We leave a check for $U_{j}f\in C_{2}$ to the reader. It is also easy to see that each $U_{j}$
is affine and $U_{j}oU_{k}=U_{k}oU_{j}$ . So by Theorem 7.1, there exists $g\in C_{3}$ such that
$\tilde{g}(a_{i})=\tilde{g}(a_{i}+a_{j})-\tilde{g}(a_{j})$ for all $i,j$ . Thus $C_{3}\neq\emptyset$ .
By $\Pi_{1}^{0}$-induction (equivalently $\Sigma_{1}^{0}$-induction), we can easily show in $RCA_{0}$ that if
$f\in C_{3}$ then
$f(na_{i})=nf(a_{i})$ for all $n\in N$ .
Hence, if $f\in C_{3}$ , then
$m \tilde{f}(\frac{n}{m}a_{i})=\tilde{f}(na_{i})=n\tilde{f}(a_{i})$ for $n,$ $m\in N$ ,
3/
75
that is,
$f(qa_{i})=q\tilde{f}(a_{i})$ for $q\in Q$ .
Therefore, any continuous function in $C_{3}$ is a bounded linear functional extending $F$ .
This completes the proof. $\square$
For the reversal of the Hahn-Banach theorem, see Metakides, Nerode and Shore
[12], and Brown and Simpson [3]. They have indeed proved that the Hahn-Banach
theorem and $WKL_{0}$ are equivalent to each other over $RCA_{0}$ .
References
[1] O. Aberth, Computable anaZysis, McGraw-Hill, New York, 1980.
[2] M. J. Beeson, Foundations of constructive analysis, Springer-Verlag, 1985.
[3] D. K. Brown and S. G. Simpson, Which set existence axioms are needed to prove
the separable Hahn-Banach theorem?, Anal. Pure Appl. Logic 31(1986), 123-
144.
[4] J. Dugundji and A. Granas, Fixed point theory, PWN-Polish Scientific Publish-
ers, Warzawa, 1982.
[5] K. Fan, A generalization of Tychonoff’s fixed point theorem, Math. Ann. 142(1961),
305-310.
[6] H. M. Friedman, Some systems of second order arithmetic and their use, in Proc.
Internat. Congress of Mathematicians (Vancouver, 1974), vol. 1, (Canadian
math. congress, 1975), 235-242.
31
76
[7] H. M. Friedman, Systems of second order arithmetic with restricted induction $I$,
$\Pi$ (abstracts), J. Symb. Logic 41(1976), 557-559.
[8] H. M. Friedman, S. G. Simpson, and R. L. Smith, Countable algebra and set
existence axioms, Annal. Pure Appl. Logic 25(1983), 141-181.
[9] D. Gale, The game of $Hex$ and the Brouwer fixed point theorem, Amer. Math.
Monthly 86(1979), 818-827.
[10] N. Hirano, H. Komiya and W. Takahashi, A generalization of the Hahn-Banach
theorem, J. Math. Anal. Appl. 88(1982), 333-340.
[11] S. Kakutani, Two fixed point theorems concerning bicompact convex sets, Proc.
Imp. Acad. Tokyo 14(1938), 242-245.
[12] G. Metakides, A. Nerode and R. A. Shore, $Rec$ursive limits on the Hahn-Banach
theorem, Cont. Math. 39(1985), 85-91.
[13] S. G. Simpson, Subsystems of second order arithmetic, forthcoming.
[14] S. G. Simpson, Which set existence axioms are needed to prove the Cauchy/Peano
theorem for ordinary differential equations?, J. Symb. Logic 49(1984), 783-802.
[15] S. G. Simpson, Reverse mathematics, Proc. Symp. Pure Math. 42(1985), 461-
471.
[16] S. G. Simpson, Subsystems of $Z_{2}$ and reverse mathematics, appendix to: G.
Takeuti, Proof Theory, North-Holland, Amsterdam, 1986, second edition.
[17] D. R. Smart, Fixed point theorems, Cambridge Univ. Press, 1974.
33
77
[18] W. Takahashi, Fixed point, minimax, and Hahn-Banach theorems, Proc. Symp.
Pure Math. 45(1986), part 2, 419-427.
$3+$